Resolving single cells in heavily clustered Nissl-stained images for the analysis of brain cytoarchitecture

The analysis of the cytoarchitecture of a tissue is of great importance for the understanding of development, behavior and disease. This is also true when analyzing tissue specimens of the brain, for analyzing cells morphology and their spatial organization. To this end, on the Nissl-stained each single cells present in the sample needs to be detected, classified according to its morphology and position. The dimension of typical histological images and the sheer numbers of cells present make the task impossible to be carried out manually. Additionally, the presence of background and staining heterogeneity, clutter, heavily clustered cells, and variability in shape and appearance of cells, makes the task difficult also for automatic methods. We present a method that building on the tentative detection obtained by local thresholding and radial symmetry transform, represent each cell cluster as a sparse mixture of gaussians. We show that the proposed method performs well both in terms of precision and recall, obtaining a F₁-score of 0.87 on Nissl-stained images of the cerebellum.